A multi-intelligent terminal collaborative decision-making method and device

Abstract

The present invention relates to a multi-intelligent terminal collaborative decision-making method and device thereof, and belongs to the field of multi-terminal collaborative edge computing. The method comprises: first, initializing the model parameters of the deep network model and the autoencoder VAE network carried by each intelligent terminal participating in the decision, and using its own local privacy data pool for model training; secondly, using differential privacy to encrypt the data, and uploading it to the cloud data sharing pool using Hash index tags; then, using the semantic cache sampling technology to extract task-specific data, and distribute it to each terminal task-specific data pool; finally, using a trusted integration algorithm to train the local model, so as to continuously interact to achieve multi-terminal collaborative decision-making. The present invention not only ensures the data transmission efficiency and privacy security, but also improves the credibility of multi-terminal decision-making, and can be applied to multi-intelligent terminal collaborative decision-making tasks with different parameters and hardware heterogeneity.

Description

Technical Field

[0001] The present invention relates to a multi-intelligent terminal collaborative decision-making method and a device thereof, and belongs to the field of multi-terminal collaborative edge computing. Background Art

[0002] With the rapid increase in the number of IoT smart terminal devices, edge intelligence has emerged. By deploying AI model training and reasoning on smart edge terminals, it has been widely used in the fields of industrial IoT, smart agriculture, smart healthcare, and unmanned driving. Traditional edge intelligence usually uploads local data directly to the cloud and trains models in a centralized manner. Although it has achieved good results, it has destroyed local data privacy. Therefore, in the context of privacy protection and data security, it is necessary not only to decentralize data privacy protection, but also to provide a fast, secure and reliable multi-terminal collaborative decision-making method to improve the accuracy and credibility of local models, which undoubtedly brings huge challenges to multi-terminal collaboration.

[0003] In related technologies, in order to achieve decentralized training, the federated learning architecture is currently used for model training. The representative algorithm is FedAvg, including a variety of variant algorithms such as FedProx, FedNova, SCAFFOLD, etc. The main idea is to train the model locally without uploading data, and upload the trained local parameter model to the central server for global model aggregation to complete the model training task in a collaborative way. However, to achieve small and convenient terminal equipment collaborative training, there are still the following deficiencies:

[0004] (1) Model heterogeneity. Different user behaviors and device functions require the deployment of devices with different parameters and heterogeneous hardware specifications, making the models deployed on each terminal heterogeneous. The traditional deployment method of federated learning, which downloads the same model from the cloud, is not applicable.

[0005] (2) Speed ​​of interaction. Due to the limitation of the amount of data in the device-cloud interaction in the mobile edge network, the traditional federated learning method of uploading all model parameters leads to low interaction efficiency.

[0006] (3) Data security. Due to the limitations of local terminal data privacy protection, the traditional method of uploading model parameters for federated learning does not meet the differential privacy constraints, resulting in privacy leakage;

[0007] (4) Credibility of model decisions. Due to the multi-source heterogeneity and distribution heterogeneity of different terminal devices, the global average aggregation used in traditional federated learning is easily affected by environmental changes and device differences, and lacks credible evidence support, making the aggregated model unreliable.

[0008] In response to the above problems, the present invention discloses a safe and reliable multi-intelligent terminal collaborative decision-making method and device, which not only ensures the data transmission efficiency and privacy security, but also improves the credibility of multi-terminal collaborative decision-making, and can be applied to multi-intelligent terminal collaborative decision-making tasks in complex scenarios. Summary of the invention

[0009] The technical problem to be solved by the present invention is to provide a multi-intelligent terminal collaborative decision-making method and a device thereof, which solves the problems in the related technology that the multi-terminal collaborative decision-making method cannot adapt to model heterogeneity, privacy leakage and low interaction efficiency, ensures data transmission efficiency and privacy security, and improves the credibility of multi-terminal collaborative decision-making.

[0010] The technical solution of the present invention is: a multi-intelligent terminal collaborative decision-making method, the specific steps are:

[0011] S1: Initialize the model parameters of the deep network and variational autoencoder (VAE) network carried by each smart terminal involved in decision-making, and use its own local privacy data pool for model training;

[0012] S2: Based on the trained local models of each smart terminal, data is encrypted using differential privacy and uploaded to the cloud data sharing pool using hash index tags;

[0013] S3: Based on the uploaded cloud data, extract task-specific data using a semantic cache sampling technique and distribute it to the task-specific data pool of each smart terminal;

[0014] S4: Based on the data from the local privacy data pool of each smart terminal and the task-specific data pool downloaded from the cloud, a trusted integration algorithm is used to perform local model training, so that continuous interaction can be achieved to achieve collaborative decision-making among multiple smart terminals.

[0015] Optionally, S1 initializes model parameters of the deep network model and autoencoder VAE network carried by each smart terminal participating in the decision-making, and uses its own local privacy data pool for model training, and the steps include:

[0016] S1.1: Initialize the pre-trained deep network installed on each smart terminal ,in represents the network coding part, Represents the task prediction part;

[0017] S1.2: Initialize the autoencoder VAE network installed on each smart terminal ;

[0018] S1.3: Perform model training based on the local privacy data pool. Given Smart terminals involved in decision-making, each with its own private privacy data ,in Indicates The amount of privacy data of smart terminals, represents a data sample, represents the sample label, Characterization of intelligent terminal devices and specific prediction tasks ,in C Indicates the number of categories of training tasks, c Represents a specific class.

[0019] Optionally, the S2 uses differential privacy to encrypt data based on the trained local models of each smart terminal, and uploads the data to a cloud data sharing pool using a Hash index tag. The steps include:

[0020] S2.1: Use differential privacy to encrypt data, where the differential privacy data encryption is a VAE representation decoupled DP streamlined data model based on mutual information optimization;

[0021] S2.2: Use Hash index tags to upload to the cloud data sharing pool, where the Hash index tags are and Two different ways of combining indexes.

[0022] Optionally, the S3 extracts task-specific data based on the uploaded cloud shared data using a semantic cache sampling technique and distributes the extracted task-specific data to the task-specific data pool of each smart terminal, and the steps include:

[0023] S3.1: Extract task-specific data using a semantic cache sampling technique, where the semantic cache sampling technique is Weight Controlled Data sampling;

[0024] S3.2: Distribute the sampling task special data obtained in S3.1 to the task special data pool of each intelligent terminal.

[0025] Optionally, S4 uses a trusted integration algorithm to perform local model training based on the data in the local privacy data pool of each smart terminal and the task-specific data pool downloaded from the cloud, so as to continuously interact to achieve collaborative decision-making among multiple smart terminals, and the steps include:

[0026] S4.1: Establish training data, which is local private data and task-specific data downloaded from the cloud The union of ;

[0027] S4.2: Perform local model training using a trusted integration algorithm based on the training data, where the trusted integration algorithm is a KL regularized optimization algorithm supported by DS evidence;

[0028] S4.3: Based on multiple data interactions between each participating smart terminal and the cloud, the above steps are repeated to achieve multi-terminal collaborative decision-making.

[0029] In addition, the present application also proposes a multi-intelligent terminal collaborative decision-making device, the multi-intelligent terminal collaborative decision-making device comprising:

[0030] (1) Deep model training module for each smart terminal, used for the deep network carried by each terminal Model training;

[0031] (2) Each intelligent terminal VAE training module is used to train the autoencoder VAE network ;

[0032] (3) Each smart terminal data upload module is used to upload data marked with differential privacy encryption and hash index to the cloud task special data pool;

[0033] (4) A trusted integrated training module for each intelligent terminal, used to optimize the model using DS evidence fusion KL loss;

[0034] (5) Cloud shared data receiving module, used to receive and store shared data uploaded by each smart terminal;

[0035] (6) Cloud-based semantic cache sampling module, used to extract task-specific data and distribute it to the task-specific data pool of each smart terminal.

[0036] In addition, the present application also proposes a multi-intelligent terminal collaborative decision-making device, including a memory, a processor, and a safe and reliable multi-intelligent terminal collaborative decision-making method program stored in the memory and runnable on the processor, and the processor executes the safe and reliable multi-intelligent terminal collaborative decision-making method program to implement the steps of the multi-intelligent terminal collaborative decision-making method as described above.

[0037] In addition, the present application also proposes a computer-readable storage medium, on which a safe and reliable multi-intelligent terminal collaborative decision-making method program is stored. When the safe and reliable multi-intelligent terminal collaborative decision-making method program is executed by the processor, the steps of the safe and reliable multi-intelligent terminal collaborative decision-making method as described above are implemented.

[0038] The beneficial effects of the present invention are:

[0039] (1) Each terminal device supports heterogeneous network deployment, which is more powerful and universal than the existing methods;

[0040] (2) The data uploaded by each terminal is marked with a hash index, which improves the efficiency of searching for similar data;

[0041] (3) Based on the mutual information optimized VAE representation decoupling DP model, each terminal device only uploads a simplified part of the feature data and satisfies differential privacy protection. It is not necessary to upload all model parameters. The cloud uses semantic cache sampling and only distributes the simplified part of the data that is special to each terminal task, thereby improving the interaction efficiency between the terminal and the cloud.

[0042] (4) The KL regularized optimization algorithm supported by the above evidence can effectively improve the credibility of task decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of the steps of the present invention;

[0044] Figure 2 This is a schematic diagram of the overall process of Example 1 of the present invention;

[0045] Figure 3 This is a schematic diagram of a VAE representation decoupled DP simplified data model based on mutual information optimization according to Example 1 of the present invention;

[0046] Figure 4 This is a schematic diagram of a cloud-based semantic cache sampling structure according to Embodiment 1 of the present invention;

[0047] Figure 5 A schematic diagram of a KL regularization optimization algorithm supported by evidence of Example 1 of the present invention;

[0048] Figure 6 This is a schematic diagram of the hardware structure involved in Example 1 of the present invention. DETAILED DESCRIPTION

[0049] In related technologies, in order to achieve decentralized training, the federated learning architecture is currently used for model training. The representative algorithm is FedAvg, including a variety of variant algorithms such as FedProx, FedNova, SCAFFOLD, etc. The main idea is to train the model locally without uploading data, and upload the trained local parameter model to the central server for global model aggregation to complete the model training task in a collaborative way. However, to achieve small and convenient terminal equipment collaborative training, there are still the following deficiencies:

[0050] (1) Model heterogeneity. Different user behaviors and device functions require the deployment of devices with different parameters and heterogeneous hardware specifications, making the models deployed on each terminal heterogeneous. The traditional deployment method of federated learning, which downloads the same model from the cloud, is not applicable.

[0051] (2) Speed ​​of interaction. Due to the limitation of the amount of data in the device-cloud interaction in the mobile edge network, the traditional federated learning method of uploading all model parameters leads to low interaction efficiency.

[0052] (3) Data security. Due to the limitations of local terminal data privacy protection, the traditional method of uploading model parameters for federated learning does not meet the differential privacy constraints, resulting in privacy leakage;

[0053] (4) Credibility of model decisions. Due to the multi-source heterogeneity and distribution heterogeneity of different terminal devices, the global average aggregation used in traditional federated learning is easily affected by environmental changes and device differences, and lacks credible evidence support, making the aggregated model unreliable.

[0054] In response to the above problems, the present invention discloses a safe and reliable multi-intelligent terminal collaborative decision-making method and device, which not only ensures the data transmission efficiency and privacy security, but also improves the credibility of multi-terminal collaborative decision-making, and can be applied to multi-intelligent terminal collaborative decision-making tasks in complex scenarios.

[0055] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation methods.

[0056] Example 1: Figure 1 As shown, a multi-intelligent terminal collaborative decision-making method, the specific steps are:

[0057] S1: Initialize the model parameters of the deep network and variational autoencoder (VAE) network carried by each smart terminal involved in decision-making, and use its own local privacy data pool for model training;

[0058] S2: Based on the trained local models of each smart terminal, data is encrypted using differential privacy and uploaded to the cloud data sharing pool using hash index tags;

[0059] S3: Based on the uploaded cloud data, extract task-specific data using a semantic cache sampling technique and distribute it to the task-specific data pool of each smart terminal;

[0060] S4: Based on the data from the local privacy data pool of each smart terminal and the task-specific data pool downloaded from the cloud, a trusted integration algorithm is used to perform local model training, so that continuous interaction can be achieved to achieve collaborative decision-making among multiple smart terminals.

[0061] like Figure 2 As shown, the specific implementation process of the method includes four steps: S1: local privacy data training; S2: data encryption and tag upload; S3: cloud data sharing semantic cache sampling; S4: trusted integration model training. The following is a detailed description of each step:

[0062] S1: Specific implementation steps for local privacy data training:

[0063] S1.1: Initialize the pre-trained deep network installed on each smart terminal ,in represents the network coding part, Represents the task prediction part;

[0064] S1.2: Initialize the autoencoder VAE network installed on each smart terminal ;

[0065] S1.3: Perform model training based on the local privacy data pool. Given Smart terminals involved in decision-making, each with its own private privacy data ,in Indicates The amount of privacy data of smart terminals, represents a data sample, represents the sample label, Characterization of intelligent terminal devices and specific prediction tasks ,in C Indicates the number of categories of training tasks, c Represents a specific class.

[0066] S2: Data encryption and tag upload specific implementation steps:

[0067] S2.1: Use differential privacy to encrypt data. The differential privacy data encryption is a VAE representation decoupled DP streamlined data model based on mutual information optimization. Specifically:

[0068] like Figure 3 As shown, it is a schematic diagram of VAE representation decoupling DP simplified data model based on mutual information optimization. Characterization of intelligent terminal devices , based on the autoencoder VAE network, the encoding features can be obtained In order to simplify the uploaded data parameters, based on the information bottleneck theory, this patent only uploads sensitive feature information and retains robust feature information for local model training. The optimization formula is as follows:

[0069] (1)

[0070] The simplified features can be obtained through formula (1) In order to meet the differential privacy protection conditions, this patent adds noise, such as Gaussian noise into the simplified feature , thereby obtaining differential privacy protection features ,in It obeys the standard normal distribution and satisfies the following conditions:

[0071] (2)

[0072] right , a random mechanism M provides - Differential privacy (DP) is a privacy policy if and only if any two adjacent datasets and Different on a single instance, where Represents the probability of failure.

[0073] S2.2: Use Hash index tags to upload to the cloud data sharing pool, where the Hash index tags are and There are two different ways of joint indexing, specifically:

[0074] Based on the predicted probability of each smart terminal class obtained in S1.3 The simplified features based on differential privacy protection obtained in S2.1 , use Hash index to mark, Indicates Smart terminal samples The index value of Indicates Smart Terminal The semantic weight of is described as follows:

[0075]

[0076] (3)

[0077] For each smart terminal index mark and Upload to the cloud shared data receiving module.

[0078] S3: Specific implementation steps for cloud data sharing semantic cache sampling:

[0079] S3.1: Extract task-specific data using a semantic cache sampling technique, where the semantic cache sampling technique is Weight Controlled Data sampling, specifically:

[0080] like Figure 4 As shown, it is a schematic diagram of the cloud-based semantic cache sampling structure. and Two different Hash indexes, Recorded the Smart terminal samples Simplify the protection feature index value, Recorded the Smart Terminal The semantic weight of Class semantic weights can be controlled Sampling is performed to reduce the amount of data downloaded by each terminal, which is described as:

[0081] (4)

[0082] in Represents the dataset Probability Randomly extract data.

[0083] S3.2: The sampled data obtained based on S3.1 is distributed to the special data pool of each intelligent terminal task. The sampled data obtained based on formula (4) is distributed to the special data pool of each intelligent terminal task for the next stage of model trusted integration training.

[0084] S4: Specific implementation steps for trusted integration model training:

[0085] S4.1: Establish training data, which is local private data and task-specific data downloaded from the cloud The union of ;

[0086] S4.2: Based on the training data, a trusted integration algorithm is used to perform local model training to improve the credibility of task decision making. The trusted integration algorithm is a KL regularized optimization algorithm supported by DS evidence. Specifically:

[0087] like Figure 5 As shown, a schematic diagram of the KL regular optimization algorithm supported by evidence is shown. Based on the data set obtained in S4.1 and , given Data ,right Classification task, based on DS evidence theory available non-negative values, described as:

[0088] (5)

[0089] in represents the uncertainty of the current data set, Indicates the current data The probability of the class. Given Evidence , the parameterized Dirichlet distribution is expressed as ,in According to formula (5), we can get:

[0090] (6)

[0091] in Denotes Dirichlet's length. Given two data sets, , , the final fusion result of the two data can be defined as , as follows:

[0092] (7)

[0093] in Measured and The number of probability value conflicts. In the actual training process, instead of using the softmax function, we only need to use the activation function Relu to ensure that the network outputs a non-negative value to obtain the evidence vector , based on Dirichlet distribution, the KL regularized optimization method supported by evidence can be described as:

[0094] (8)

[0095] in represents the digamma function, , the evidence support using formula (8) is more suitable for the fusion of multiple decisions and can effectively improve the credibility of task decisions.

[0096] S4.3: Based on multiple data interactions between each participating smart terminal and the cloud, the above steps are repeated to achieve multi-terminal collaborative decision-making.

[0097] This application also proposes a multi-intelligent terminal collaborative decision-making device, such as Figure 6 As shown, Figure 6 A schematic diagram of the structure of a device for running a program of a secure and reliable multi-terminal collaborative decision-making method in a hardware operating environment involved in an embodiment of the present application.

[0098] like Figure 6As shown, the multi-intelligent terminal collaborative decision-making device includes: a processor 1001, such as a central processing unit CPU, a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. Among them, the communication bus 1002 is used to realize the connection and communication between each component module, the user interface 1003 may include a display screen, an input unit such as a keyboard, etc., and the user interface 1003 may optionally include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed random access memory RAM, or a stable non-volatile memory NVM, such as a disk storage. The memory 1005 may optionally be a storage device independent of the aforementioned processor 1001.

[0099] Those skilled in the art will understand that Figure 6 The structure shown in does not constitute a limitation on a multi-intelligent terminal collaborative decision-making device, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0100] Optionally, the memory 1005 is electrically connected to the processor 1001, and the processor 1001 can be used to control the operation of the memory 1005, and can also read the data in the memory 1005 to implement a multi-intelligent terminal collaborative decision-making method.

[0101] Alternatively, if Figure 6 As shown, the memory 1005 as a storage medium may include an operating system, a data storage module, a network communication module, a user interface module, and a secure and reliable multi-intelligent terminal collaborative decision-making method program.

[0102] Optionally, in Figure 6 In the multi-intelligent terminal collaborative decision-making device shown, the network interface 1004 is mainly used for data communication with other devices; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the multi-intelligent terminal collaborative decision-making device can be set in a multi-intelligent terminal collaborative decision-making device.

[0103] like Figure 6 As shown, the multi-intelligent terminal collaborative decision-making device calls the secure and reliable multi-intelligent terminal collaborative decision-making method program stored in the memory 1005 through the processor 1001, and executes the following Figure 2 As shown, there are four related steps of a multi-intelligent terminal collaborative decision-making method: S1: local privacy data training; S2: data encryption and labeling upload; S3: cloud data sharing semantic cache sampling; S4: trusted integration model training.

[0104] In addition, the present application also proposes a computer-readable storage medium, on which a safe and reliable multi-intelligent terminal collaborative decision-making method program is stored. When the safe and reliable multi-intelligent terminal collaborative decision-making method program is executed by the processor, the steps of the multi-intelligent terminal collaborative decision-making method as described above are implemented.

[0105] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] It should be noted that in the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The present application may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0107] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0108] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A multi-intelligent terminal collaborative decision-making method, characterized in that: The multi-intelligent terminal collaborative decision-making method includes: S1: Initialize the model parameters of the deep network and autoencoder VAE network carried by each smart terminal involved in decision-making, and use its own local privacy data pool for model training; S2: Based on the trained local models of each smart terminal, data is encrypted using differential privacy and uploaded to the cloud data sharing pool using hash index tags; S3: Based on the uploaded cloud data, extract task-specific data using a semantic cache sampling technique and distribute it to the task-specific data pool of each smart terminal; S4: Based on the data in the local privacy data pool of each smart terminal and the task-specific data pool downloaded from the cloud, a trusted integration algorithm is used to perform local model training, so that continuous interaction is performed to achieve collaborative decision-making among multiple smart terminals; S1 initializes the model parameters of the deep network model and autoencoder VAE network carried by each smart terminal involved in the decision-making, and uses its own local privacy data pool for model training. The steps include: S1.1: Initialize the pre-trained deep network installed on each smart terminal ,in represents the network coding part, Represents the task prediction part; S1.2: Initialize the autoencoder VAE network installed on each smart terminal ; S1.3: Perform model training based on the local privacy data pool. Given Smart terminals involved in decision-making, each with its own private privacy data ,in Indicates The amount of privacy data of smart terminals, represents a data sample, represents the sample label, Characterization of intelligent terminal devices and specific prediction tasks ,in C Indicates the number of categories of training tasks, c Represents a specific class; The S2 uses differential privacy to encrypt data based on the trained local models of each smart terminal, and uploads it to the cloud data sharing pool using a hash index tag. The steps include: S2.1: Use differential privacy to encrypt data. The differential privacy data encryption is a VAE representation decoupled DP streamlined data model based on mutual information optimization. Specifically: Based on the S1.3 Characterization of intelligent terminal devices , based on the autoencoder VAE network, the encoding features can be obtained In order to simplify the uploaded data parameters, based on the information bottleneck theory, only sensitive feature information is uploaded and robust feature information is retained for local model training. The optimization formula is as follows: (1) The simplified features can be obtained through formula (1) In order to meet the differential privacy protection conditions, Gaussian noise is used to enter the simplified feature , thereby obtaining differential privacy protection features ,in It obeys the standard normal distribution and satisfies the following conditions: (2) right , a random mechanism M provides - Differential privacy DP if and only if any two adjacent data sets and Different on a single instance, where represents the probability of failure; S2.2: Use Hash index tags to upload to the cloud data sharing pool, where the Hash index tags are and There are two different ways of joint indexing, specifically: Based on the predicted probability of each smart terminal class obtained in S1.3 The simplified features based on differential privacy protection obtained in S2.1 , use Hash index to mark, Indicates Smart terminal samples The index value of Indicates Smart Terminal The semantic weight of is described as follows: ； (3) For each smart terminal index mark and Upload to the cloud shared data receiving module; The S3 extracts task-specific data based on the uploaded cloud shared data using a semantic cache sampling technique and distributes it to the task-specific data pool of each smart terminal. The steps include: S3.1: Extract task-specific data using a semantic cache sampling technique, where the semantic cache sampling technique is Weight Controlled Data sampling; S3.2: Distribute the sampling task special data obtained in S3.1 to the task special data pool of each intelligent terminal; The S4 uses a trusted integration algorithm to perform local model training based on the data in the local privacy data pool of each smart terminal and the task-specific data pool downloaded from the cloud, so as to continuously interact to achieve collaborative decision-making among multiple smart terminals. The steps include: S4.1: Establish training data, which is local private data and task-specific data downloaded from the cloud The union of ; S4.2: Perform local model training using a trusted integration algorithm based on the training data, where the trusted integration algorithm is a KL regularized optimization algorithm supported by DS evidence; S4.3: Based on multiple data interactions between each smart terminal and the cloud, the above steps are repeated to achieve multi-terminal collaborative decision-making.

2. A device for implementing the multi-intelligent terminal collaborative decision-making method as claimed in claim 1, characterized in that: The multi-intelligent terminal collaborative decision-making device includes: Deep model training modules for each smart terminal, used for the deep network carried by each terminal Model training; Each intelligent terminal VAE training module is used to train the autoencoder VAE network ; Each smart terminal data upload module is used to upload data marked with differential privacy encryption and hash index to the cloud task special data pool; The trusted integrated training module of each intelligent terminal is used to optimize the model using DS evidence fusion KL loss; The cloud shared data receiving module is used to receive and store the shared data uploaded by each smart terminal; The cloud-based semantic cache sampling module is used to extract task-specific data and distribute it to the task-specific data pool of each smart terminal.

3. A multi-intelligent terminal collaborative decision-making device, characterized in that: It includes a memory, a processor, and a safe and reliable multi-intelligent terminal collaborative decision-making method program stored in the memory and executable on the processor, wherein the processor executes the safe and reliable multi-intelligent terminal collaborative decision-making method program to implement the steps of the multi-intelligent terminal collaborative decision-making method as described in claim 1.

4. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a safe and reliable multi-intelligent terminal collaborative decision-making method program, and when the safe and reliable multi-intelligent terminal collaborative decision-making method program is executed by the processor, the steps of the multi-intelligent terminal collaborative decision-making method as described in claim 1 are implemented.

Images